An optimal scheduling method for building energy system integrating solar energy based on adaptive hybrid mechanism model and learning-based model predictive control

Building energy system integrating solar energy as a new form of sustainable energy use, reduces energy consumption and provides great flexibility for energy saving and emission reduction; however, the stochasticity and fluctuation of solar energy supply makes it difficult to utilize efficiently and reduce the economy of building energy system. Optimal scheduling can fully coordinate the output of each equipment unit, effectively improve the energy utilization rate and further enhance the economic operation; however, optimal scheduling method depends on the accurate system model, and the uncertainty and lack of accuracy hinder the effective implementation of the optimal scheduling method. This paper presents an optimal scheduling method based on adaptive hybrid mechanism model (AHMM) and learning-based model predictive control (LMPC). Firstly, an AHMM is constructed to profile the actual output of the system more accurately to help optimal scheduling methods to be implemented more effectively, in which mechanism model is used to explain the theoretical output characteristics of the system, and the data-driven model is introduced to describe and quantify the uncertainty of the energy coupling. Secondly, an optimal scheduling based on LMPC is proposed, wherein, LMPC with dynamic optimization characteristics is designed to address the economic performance degradation caused by fluctuations in solar energy supply, and the prediction horizon of LMPC is updated by an online learning mechanism to improve the economic performance and computational efficiency of optimal scheduling. Finally, a typical building energy system based on simulation platform is used as the experimental object. The results illustrate that the proposed AHMM has more accurate output and smaller prediction error, and the proposed optimal scheduling method improves the cost saving and computational efficiency by about 21.09% and 17.28% respectively. The proposed method can quantify the uncertainty of the model, better profile the system response, and have application potential and advantage under the situation of the declining computer efficiency.